xref: /linux/mm/z3fold.c (revision b68fc09be48edbc47de1a0f3d42ef8adf6c0ac55)
1 /*
2  * z3fold.c
3  *
4  * Author: Vitaly Wool <vitaly.wool@konsulko.com>
5  * Copyright (C) 2016, Sony Mobile Communications Inc.
6  *
7  * This implementation is based on zbud written by Seth Jennings.
8  *
9  * z3fold is an special purpose allocator for storing compressed pages. It
10  * can store up to three compressed pages per page which improves the
11  * compression ratio of zbud while retaining its main concepts (e. g. always
12  * storing an integral number of objects per page) and simplicity.
13  * It still has simple and deterministic reclaim properties that make it
14  * preferable to a higher density approach (with no requirement on integral
15  * number of object per page) when reclaim is used.
16  *
17  * As in zbud, pages are divided into "chunks".  The size of the chunks is
18  * fixed at compile time and is determined by NCHUNKS_ORDER below.
19  *
20  * z3fold doesn't export any API and is meant to be used via zpool API.
21  */
22 
23 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
24 
25 #include <linux/atomic.h>
26 #include <linux/sched.h>
27 #include <linux/list.h>
28 #include <linux/mm.h>
29 #include <linux/module.h>
30 #include <linux/percpu.h>
31 #include <linux/preempt.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/spinlock.h>
35 #include <linux/zpool.h>
36 
37 /*****************
38  * Structures
39 *****************/
40 struct z3fold_pool;
41 struct z3fold_ops {
42 	int (*evict)(struct z3fold_pool *pool, unsigned long handle);
43 };
44 
45 enum buddy {
46 	HEADLESS = 0,
47 	FIRST,
48 	MIDDLE,
49 	LAST,
50 	BUDDIES_MAX
51 };
52 
53 /*
54  * struct z3fold_header - z3fold page metadata occupying first chunks of each
55  *			z3fold page, except for HEADLESS pages
56  * @buddy:		links the z3fold page into the relevant list in the
57  *			pool
58  * @page_lock:		per-page lock
59  * @refcount:		reference count for the z3fold page
60  * @work:		work_struct for page layout optimization
61  * @pool:		pointer to the pool which this page belongs to
62  * @cpu:		CPU which this page "belongs" to
63  * @first_chunks:	the size of the first buddy in chunks, 0 if free
64  * @middle_chunks:	the size of the middle buddy in chunks, 0 if free
65  * @last_chunks:	the size of the last buddy in chunks, 0 if free
66  * @first_num:		the starting number (for the first handle)
67  */
68 struct z3fold_header {
69 	struct list_head buddy;
70 	spinlock_t page_lock;
71 	struct kref refcount;
72 	struct work_struct work;
73 	struct z3fold_pool *pool;
74 	short cpu;
75 	unsigned short first_chunks;
76 	unsigned short middle_chunks;
77 	unsigned short last_chunks;
78 	unsigned short start_middle;
79 	unsigned short first_num:2;
80 };
81 
82 /*
83  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
84  * adjusting internal fragmentation.  It also determines the number of
85  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
86  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
87  * in the beginning of an allocated page are occupied by z3fold header, so
88  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
89  * which shows the max number of free chunks in z3fold page, also there will
90  * be 63, or 62, respectively, freelists per pool.
91  */
92 #define NCHUNKS_ORDER	6
93 
94 #define CHUNK_SHIFT	(PAGE_SHIFT - NCHUNKS_ORDER)
95 #define CHUNK_SIZE	(1 << CHUNK_SHIFT)
96 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
97 #define ZHDR_CHUNKS	(ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
98 #define TOTAL_CHUNKS	(PAGE_SIZE >> CHUNK_SHIFT)
99 #define NCHUNKS		((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
100 
101 #define BUDDY_MASK	(0x3)
102 
103 /**
104  * struct z3fold_pool - stores metadata for each z3fold pool
105  * @name:	pool name
106  * @lock:	protects pool unbuddied/lru lists
107  * @stale_lock:	protects pool stale page list
108  * @unbuddied:	per-cpu array of lists tracking z3fold pages that contain 2-
109  *		buddies; the list each z3fold page is added to depends on
110  *		the size of its free region.
111  * @lru:	list tracking the z3fold pages in LRU order by most recently
112  *		added buddy.
113  * @stale:	list of pages marked for freeing
114  * @pages_nr:	number of z3fold pages in the pool.
115  * @ops:	pointer to a structure of user defined operations specified at
116  *		pool creation time.
117  * @compact_wq:	workqueue for page layout background optimization
118  * @release_wq:	workqueue for safe page release
119  * @work:	work_struct for safe page release
120  *
121  * This structure is allocated at pool creation time and maintains metadata
122  * pertaining to a particular z3fold pool.
123  */
124 struct z3fold_pool {
125 	const char *name;
126 	spinlock_t lock;
127 	spinlock_t stale_lock;
128 	struct list_head *unbuddied;
129 	struct list_head lru;
130 	struct list_head stale;
131 	atomic64_t pages_nr;
132 	const struct z3fold_ops *ops;
133 	struct zpool *zpool;
134 	const struct zpool_ops *zpool_ops;
135 	struct workqueue_struct *compact_wq;
136 	struct workqueue_struct *release_wq;
137 	struct work_struct work;
138 };
139 
140 /*
141  * Internal z3fold page flags
142  */
143 enum z3fold_page_flags {
144 	PAGE_HEADLESS = 0,
145 	MIDDLE_CHUNK_MAPPED,
146 	NEEDS_COMPACTING,
147 	PAGE_STALE,
148 	UNDER_RECLAIM
149 };
150 
151 /*****************
152  * Helpers
153 *****************/
154 
155 /* Converts an allocation size in bytes to size in z3fold chunks */
156 static int size_to_chunks(size_t size)
157 {
158 	return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
159 }
160 
161 #define for_each_unbuddied_list(_iter, _begin) \
162 	for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
163 
164 static void compact_page_work(struct work_struct *w);
165 
166 /* Initializes the z3fold header of a newly allocated z3fold page */
167 static struct z3fold_header *init_z3fold_page(struct page *page,
168 					struct z3fold_pool *pool)
169 {
170 	struct z3fold_header *zhdr = page_address(page);
171 
172 	INIT_LIST_HEAD(&page->lru);
173 	clear_bit(PAGE_HEADLESS, &page->private);
174 	clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
175 	clear_bit(NEEDS_COMPACTING, &page->private);
176 	clear_bit(PAGE_STALE, &page->private);
177 	clear_bit(UNDER_RECLAIM, &page->private);
178 
179 	spin_lock_init(&zhdr->page_lock);
180 	kref_init(&zhdr->refcount);
181 	zhdr->first_chunks = 0;
182 	zhdr->middle_chunks = 0;
183 	zhdr->last_chunks = 0;
184 	zhdr->first_num = 0;
185 	zhdr->start_middle = 0;
186 	zhdr->cpu = -1;
187 	zhdr->pool = pool;
188 	INIT_LIST_HEAD(&zhdr->buddy);
189 	INIT_WORK(&zhdr->work, compact_page_work);
190 	return zhdr;
191 }
192 
193 /* Resets the struct page fields and frees the page */
194 static void free_z3fold_page(struct page *page)
195 {
196 	__free_page(page);
197 }
198 
199 /* Lock a z3fold page */
200 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
201 {
202 	spin_lock(&zhdr->page_lock);
203 }
204 
205 /* Try to lock a z3fold page */
206 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
207 {
208 	return spin_trylock(&zhdr->page_lock);
209 }
210 
211 /* Unlock a z3fold page */
212 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
213 {
214 	spin_unlock(&zhdr->page_lock);
215 }
216 
217 /*
218  * Encodes the handle of a particular buddy within a z3fold page
219  * Pool lock should be held as this function accesses first_num
220  */
221 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
222 {
223 	unsigned long handle;
224 
225 	handle = (unsigned long)zhdr;
226 	if (bud != HEADLESS)
227 		handle += (bud + zhdr->first_num) & BUDDY_MASK;
228 	return handle;
229 }
230 
231 /* Returns the z3fold page where a given handle is stored */
232 static struct z3fold_header *handle_to_z3fold_header(unsigned long handle)
233 {
234 	return (struct z3fold_header *)(handle & PAGE_MASK);
235 }
236 
237 /*
238  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
239  *  but that doesn't matter. because the masking will result in the
240  *  correct buddy number.
241  */
242 static enum buddy handle_to_buddy(unsigned long handle)
243 {
244 	struct z3fold_header *zhdr = handle_to_z3fold_header(handle);
245 	return (handle - zhdr->first_num) & BUDDY_MASK;
246 }
247 
248 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
249 {
250 	struct page *page = virt_to_page(zhdr);
251 	struct z3fold_pool *pool = zhdr->pool;
252 
253 	WARN_ON(!list_empty(&zhdr->buddy));
254 	set_bit(PAGE_STALE, &page->private);
255 	clear_bit(NEEDS_COMPACTING, &page->private);
256 	spin_lock(&pool->lock);
257 	if (!list_empty(&page->lru))
258 		list_del(&page->lru);
259 	spin_unlock(&pool->lock);
260 	if (locked)
261 		z3fold_page_unlock(zhdr);
262 	spin_lock(&pool->stale_lock);
263 	list_add(&zhdr->buddy, &pool->stale);
264 	queue_work(pool->release_wq, &pool->work);
265 	spin_unlock(&pool->stale_lock);
266 }
267 
268 static void __attribute__((__unused__))
269 			release_z3fold_page(struct kref *ref)
270 {
271 	struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
272 						refcount);
273 	__release_z3fold_page(zhdr, false);
274 }
275 
276 static void release_z3fold_page_locked(struct kref *ref)
277 {
278 	struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
279 						refcount);
280 	WARN_ON(z3fold_page_trylock(zhdr));
281 	__release_z3fold_page(zhdr, true);
282 }
283 
284 static void release_z3fold_page_locked_list(struct kref *ref)
285 {
286 	struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
287 					       refcount);
288 	spin_lock(&zhdr->pool->lock);
289 	list_del_init(&zhdr->buddy);
290 	spin_unlock(&zhdr->pool->lock);
291 
292 	WARN_ON(z3fold_page_trylock(zhdr));
293 	__release_z3fold_page(zhdr, true);
294 }
295 
296 static void free_pages_work(struct work_struct *w)
297 {
298 	struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
299 
300 	spin_lock(&pool->stale_lock);
301 	while (!list_empty(&pool->stale)) {
302 		struct z3fold_header *zhdr = list_first_entry(&pool->stale,
303 						struct z3fold_header, buddy);
304 		struct page *page = virt_to_page(zhdr);
305 
306 		list_del(&zhdr->buddy);
307 		if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
308 			continue;
309 		spin_unlock(&pool->stale_lock);
310 		cancel_work_sync(&zhdr->work);
311 		free_z3fold_page(page);
312 		cond_resched();
313 		spin_lock(&pool->stale_lock);
314 	}
315 	spin_unlock(&pool->stale_lock);
316 }
317 
318 /*
319  * Returns the number of free chunks in a z3fold page.
320  * NB: can't be used with HEADLESS pages.
321  */
322 static int num_free_chunks(struct z3fold_header *zhdr)
323 {
324 	int nfree;
325 	/*
326 	 * If there is a middle object, pick up the bigger free space
327 	 * either before or after it. Otherwise just subtract the number
328 	 * of chunks occupied by the first and the last objects.
329 	 */
330 	if (zhdr->middle_chunks != 0) {
331 		int nfree_before = zhdr->first_chunks ?
332 			0 : zhdr->start_middle - ZHDR_CHUNKS;
333 		int nfree_after = zhdr->last_chunks ?
334 			0 : TOTAL_CHUNKS -
335 				(zhdr->start_middle + zhdr->middle_chunks);
336 		nfree = max(nfree_before, nfree_after);
337 	} else
338 		nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
339 	return nfree;
340 }
341 
342 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
343 				unsigned short dst_chunk)
344 {
345 	void *beg = zhdr;
346 	return memmove(beg + (dst_chunk << CHUNK_SHIFT),
347 		       beg + (zhdr->start_middle << CHUNK_SHIFT),
348 		       zhdr->middle_chunks << CHUNK_SHIFT);
349 }
350 
351 #define BIG_CHUNK_GAP	3
352 /* Has to be called with lock held */
353 static int z3fold_compact_page(struct z3fold_header *zhdr)
354 {
355 	struct page *page = virt_to_page(zhdr);
356 
357 	if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
358 		return 0; /* can't move middle chunk, it's used */
359 
360 	if (zhdr->middle_chunks == 0)
361 		return 0; /* nothing to compact */
362 
363 	if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
364 		/* move to the beginning */
365 		mchunk_memmove(zhdr, ZHDR_CHUNKS);
366 		zhdr->first_chunks = zhdr->middle_chunks;
367 		zhdr->middle_chunks = 0;
368 		zhdr->start_middle = 0;
369 		zhdr->first_num++;
370 		return 1;
371 	}
372 
373 	/*
374 	 * moving data is expensive, so let's only do that if
375 	 * there's substantial gain (at least BIG_CHUNK_GAP chunks)
376 	 */
377 	if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
378 	    zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
379 			BIG_CHUNK_GAP) {
380 		mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
381 		zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
382 		return 1;
383 	} else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
384 		   TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
385 					+ zhdr->middle_chunks) >=
386 			BIG_CHUNK_GAP) {
387 		unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
388 			zhdr->middle_chunks;
389 		mchunk_memmove(zhdr, new_start);
390 		zhdr->start_middle = new_start;
391 		return 1;
392 	}
393 
394 	return 0;
395 }
396 
397 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
398 {
399 	struct z3fold_pool *pool = zhdr->pool;
400 	struct page *page;
401 	struct list_head *unbuddied;
402 	int fchunks;
403 
404 	page = virt_to_page(zhdr);
405 	if (locked)
406 		WARN_ON(z3fold_page_trylock(zhdr));
407 	else
408 		z3fold_page_lock(zhdr);
409 	if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
410 		z3fold_page_unlock(zhdr);
411 		return;
412 	}
413 	spin_lock(&pool->lock);
414 	list_del_init(&zhdr->buddy);
415 	spin_unlock(&pool->lock);
416 
417 	if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
418 		atomic64_dec(&pool->pages_nr);
419 		return;
420 	}
421 
422 	z3fold_compact_page(zhdr);
423 	unbuddied = get_cpu_ptr(pool->unbuddied);
424 	fchunks = num_free_chunks(zhdr);
425 	if (fchunks < NCHUNKS &&
426 	    (!zhdr->first_chunks || !zhdr->middle_chunks ||
427 			!zhdr->last_chunks)) {
428 		/* the page's not completely free and it's unbuddied */
429 		spin_lock(&pool->lock);
430 		list_add(&zhdr->buddy, &unbuddied[fchunks]);
431 		spin_unlock(&pool->lock);
432 		zhdr->cpu = smp_processor_id();
433 	}
434 	put_cpu_ptr(pool->unbuddied);
435 	z3fold_page_unlock(zhdr);
436 }
437 
438 static void compact_page_work(struct work_struct *w)
439 {
440 	struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
441 						work);
442 
443 	do_compact_page(zhdr, false);
444 }
445 
446 
447 /*
448  * API Functions
449  */
450 
451 /**
452  * z3fold_create_pool() - create a new z3fold pool
453  * @name:	pool name
454  * @gfp:	gfp flags when allocating the z3fold pool structure
455  * @ops:	user-defined operations for the z3fold pool
456  *
457  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
458  * failed.
459  */
460 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
461 		const struct z3fold_ops *ops)
462 {
463 	struct z3fold_pool *pool = NULL;
464 	int i, cpu;
465 
466 	pool = kzalloc(sizeof(struct z3fold_pool), gfp);
467 	if (!pool)
468 		goto out;
469 	spin_lock_init(&pool->lock);
470 	spin_lock_init(&pool->stale_lock);
471 	pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
472 	if (!pool->unbuddied)
473 		goto out_pool;
474 	for_each_possible_cpu(cpu) {
475 		struct list_head *unbuddied =
476 				per_cpu_ptr(pool->unbuddied, cpu);
477 		for_each_unbuddied_list(i, 0)
478 			INIT_LIST_HEAD(&unbuddied[i]);
479 	}
480 	INIT_LIST_HEAD(&pool->lru);
481 	INIT_LIST_HEAD(&pool->stale);
482 	atomic64_set(&pool->pages_nr, 0);
483 	pool->name = name;
484 	pool->compact_wq = create_singlethread_workqueue(pool->name);
485 	if (!pool->compact_wq)
486 		goto out_unbuddied;
487 	pool->release_wq = create_singlethread_workqueue(pool->name);
488 	if (!pool->release_wq)
489 		goto out_wq;
490 	INIT_WORK(&pool->work, free_pages_work);
491 	pool->ops = ops;
492 	return pool;
493 
494 out_wq:
495 	destroy_workqueue(pool->compact_wq);
496 out_unbuddied:
497 	free_percpu(pool->unbuddied);
498 out_pool:
499 	kfree(pool);
500 out:
501 	return NULL;
502 }
503 
504 /**
505  * z3fold_destroy_pool() - destroys an existing z3fold pool
506  * @pool:	the z3fold pool to be destroyed
507  *
508  * The pool should be emptied before this function is called.
509  */
510 static void z3fold_destroy_pool(struct z3fold_pool *pool)
511 {
512 	destroy_workqueue(pool->release_wq);
513 	destroy_workqueue(pool->compact_wq);
514 	kfree(pool);
515 }
516 
517 /**
518  * z3fold_alloc() - allocates a region of a given size
519  * @pool:	z3fold pool from which to allocate
520  * @size:	size in bytes of the desired allocation
521  * @gfp:	gfp flags used if the pool needs to grow
522  * @handle:	handle of the new allocation
523  *
524  * This function will attempt to find a free region in the pool large enough to
525  * satisfy the allocation request.  A search of the unbuddied lists is
526  * performed first. If no suitable free region is found, then a new page is
527  * allocated and added to the pool to satisfy the request.
528  *
529  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
530  * as z3fold pool pages.
531  *
532  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
533  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
534  * a new page.
535  */
536 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
537 			unsigned long *handle)
538 {
539 	int chunks = 0, i, freechunks;
540 	struct z3fold_header *zhdr = NULL;
541 	struct page *page = NULL;
542 	enum buddy bud;
543 	bool can_sleep = gfpflags_allow_blocking(gfp);
544 
545 	if (!size || (gfp & __GFP_HIGHMEM))
546 		return -EINVAL;
547 
548 	if (size > PAGE_SIZE)
549 		return -ENOSPC;
550 
551 	if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
552 		bud = HEADLESS;
553 	else {
554 		struct list_head *unbuddied;
555 		chunks = size_to_chunks(size);
556 
557 lookup:
558 		/* First, try to find an unbuddied z3fold page. */
559 		unbuddied = get_cpu_ptr(pool->unbuddied);
560 		for_each_unbuddied_list(i, chunks) {
561 			struct list_head *l = &unbuddied[i];
562 
563 			zhdr = list_first_entry_or_null(READ_ONCE(l),
564 						struct z3fold_header, buddy);
565 
566 			if (!zhdr)
567 				continue;
568 
569 			/* Re-check under lock. */
570 			spin_lock(&pool->lock);
571 			l = &unbuddied[i];
572 			if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
573 					struct z3fold_header, buddy)) ||
574 			    !z3fold_page_trylock(zhdr)) {
575 				spin_unlock(&pool->lock);
576 				put_cpu_ptr(pool->unbuddied);
577 				goto lookup;
578 			}
579 			list_del_init(&zhdr->buddy);
580 			zhdr->cpu = -1;
581 			spin_unlock(&pool->lock);
582 
583 			page = virt_to_page(zhdr);
584 			if (test_bit(NEEDS_COMPACTING, &page->private)) {
585 				z3fold_page_unlock(zhdr);
586 				zhdr = NULL;
587 				put_cpu_ptr(pool->unbuddied);
588 				if (can_sleep)
589 					cond_resched();
590 				goto lookup;
591 			}
592 
593 			/*
594 			 * this page could not be removed from its unbuddied
595 			 * list while pool lock was held, and then we've taken
596 			 * page lock so kref_put could not be called before
597 			 * we got here, so it's safe to just call kref_get()
598 			 */
599 			kref_get(&zhdr->refcount);
600 			break;
601 		}
602 		put_cpu_ptr(pool->unbuddied);
603 
604 		if (zhdr) {
605 			if (zhdr->first_chunks == 0) {
606 				if (zhdr->middle_chunks != 0 &&
607 				    chunks >= zhdr->start_middle)
608 					bud = LAST;
609 				else
610 					bud = FIRST;
611 			} else if (zhdr->last_chunks == 0)
612 				bud = LAST;
613 			else if (zhdr->middle_chunks == 0)
614 				bud = MIDDLE;
615 			else {
616 				if (kref_put(&zhdr->refcount,
617 					     release_z3fold_page_locked))
618 					atomic64_dec(&pool->pages_nr);
619 				else
620 					z3fold_page_unlock(zhdr);
621 				pr_err("No free chunks in unbuddied\n");
622 				WARN_ON(1);
623 				goto lookup;
624 			}
625 			goto found;
626 		}
627 		bud = FIRST;
628 	}
629 
630 	page = NULL;
631 	if (can_sleep) {
632 		spin_lock(&pool->stale_lock);
633 		zhdr = list_first_entry_or_null(&pool->stale,
634 						struct z3fold_header, buddy);
635 		/*
636 		 * Before allocating a page, let's see if we can take one from
637 		 * the stale pages list. cancel_work_sync() can sleep so we
638 		 * limit this case to the contexts where we can sleep
639 		 */
640 		if (zhdr) {
641 			list_del(&zhdr->buddy);
642 			spin_unlock(&pool->stale_lock);
643 			cancel_work_sync(&zhdr->work);
644 			page = virt_to_page(zhdr);
645 		} else {
646 			spin_unlock(&pool->stale_lock);
647 		}
648 	}
649 	if (!page)
650 		page = alloc_page(gfp);
651 
652 	if (!page)
653 		return -ENOMEM;
654 
655 	atomic64_inc(&pool->pages_nr);
656 	zhdr = init_z3fold_page(page, pool);
657 
658 	if (bud == HEADLESS) {
659 		set_bit(PAGE_HEADLESS, &page->private);
660 		goto headless;
661 	}
662 	z3fold_page_lock(zhdr);
663 
664 found:
665 	if (bud == FIRST)
666 		zhdr->first_chunks = chunks;
667 	else if (bud == LAST)
668 		zhdr->last_chunks = chunks;
669 	else {
670 		zhdr->middle_chunks = chunks;
671 		zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
672 	}
673 
674 	if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
675 			zhdr->middle_chunks == 0) {
676 		struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
677 
678 		/* Add to unbuddied list */
679 		freechunks = num_free_chunks(zhdr);
680 		spin_lock(&pool->lock);
681 		list_add(&zhdr->buddy, &unbuddied[freechunks]);
682 		spin_unlock(&pool->lock);
683 		zhdr->cpu = smp_processor_id();
684 		put_cpu_ptr(pool->unbuddied);
685 	}
686 
687 headless:
688 	spin_lock(&pool->lock);
689 	/* Add/move z3fold page to beginning of LRU */
690 	if (!list_empty(&page->lru))
691 		list_del(&page->lru);
692 
693 	list_add(&page->lru, &pool->lru);
694 
695 	*handle = encode_handle(zhdr, bud);
696 	spin_unlock(&pool->lock);
697 	if (bud != HEADLESS)
698 		z3fold_page_unlock(zhdr);
699 
700 	return 0;
701 }
702 
703 /**
704  * z3fold_free() - frees the allocation associated with the given handle
705  * @pool:	pool in which the allocation resided
706  * @handle:	handle associated with the allocation returned by z3fold_alloc()
707  *
708  * In the case that the z3fold page in which the allocation resides is under
709  * reclaim, as indicated by the PG_reclaim flag being set, this function
710  * only sets the first|last_chunks to 0.  The page is actually freed
711  * once both buddies are evicted (see z3fold_reclaim_page() below).
712  */
713 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
714 {
715 	struct z3fold_header *zhdr;
716 	struct page *page;
717 	enum buddy bud;
718 
719 	zhdr = handle_to_z3fold_header(handle);
720 	page = virt_to_page(zhdr);
721 
722 	if (test_bit(PAGE_HEADLESS, &page->private)) {
723 		/* HEADLESS page stored */
724 		bud = HEADLESS;
725 	} else {
726 		z3fold_page_lock(zhdr);
727 		bud = handle_to_buddy(handle);
728 
729 		switch (bud) {
730 		case FIRST:
731 			zhdr->first_chunks = 0;
732 			break;
733 		case MIDDLE:
734 			zhdr->middle_chunks = 0;
735 			zhdr->start_middle = 0;
736 			break;
737 		case LAST:
738 			zhdr->last_chunks = 0;
739 			break;
740 		default:
741 			pr_err("%s: unknown bud %d\n", __func__, bud);
742 			WARN_ON(1);
743 			z3fold_page_unlock(zhdr);
744 			return;
745 		}
746 	}
747 
748 	if (bud == HEADLESS) {
749 		spin_lock(&pool->lock);
750 		list_del(&page->lru);
751 		spin_unlock(&pool->lock);
752 		free_z3fold_page(page);
753 		atomic64_dec(&pool->pages_nr);
754 		return;
755 	}
756 
757 	if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
758 		atomic64_dec(&pool->pages_nr);
759 		return;
760 	}
761 	if (test_bit(UNDER_RECLAIM, &page->private)) {
762 		z3fold_page_unlock(zhdr);
763 		return;
764 	}
765 	if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
766 		z3fold_page_unlock(zhdr);
767 		return;
768 	}
769 	if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
770 		spin_lock(&pool->lock);
771 		list_del_init(&zhdr->buddy);
772 		spin_unlock(&pool->lock);
773 		zhdr->cpu = -1;
774 		kref_get(&zhdr->refcount);
775 		do_compact_page(zhdr, true);
776 		return;
777 	}
778 	kref_get(&zhdr->refcount);
779 	queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
780 	z3fold_page_unlock(zhdr);
781 }
782 
783 /**
784  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
785  * @pool:	pool from which a page will attempt to be evicted
786  * @retries:	number of pages on the LRU list for which eviction will
787  *		be attempted before failing
788  *
789  * z3fold reclaim is different from normal system reclaim in that it is done
790  * from the bottom, up. This is because only the bottom layer, z3fold, has
791  * information on how the allocations are organized within each z3fold page.
792  * This has the potential to create interesting locking situations between
793  * z3fold and the user, however.
794  *
795  * To avoid these, this is how z3fold_reclaim_page() should be called:
796  *
797  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
798  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
799  * call the user-defined eviction handler with the pool and handle as
800  * arguments.
801  *
802  * If the handle can not be evicted, the eviction handler should return
803  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
804  * appropriate list and try the next z3fold page on the LRU up to
805  * a user defined number of retries.
806  *
807  * If the handle is successfully evicted, the eviction handler should
808  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
809  * contains logic to delay freeing the page if the page is under reclaim,
810  * as indicated by the setting of the PG_reclaim flag on the underlying page.
811  *
812  * If all buddies in the z3fold page are successfully evicted, then the
813  * z3fold page can be freed.
814  *
815  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
816  * no pages to evict or an eviction handler is not registered, -EAGAIN if
817  * the retry limit was hit.
818  */
819 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
820 {
821 	int i, ret = 0;
822 	struct z3fold_header *zhdr = NULL;
823 	struct page *page = NULL;
824 	struct list_head *pos;
825 	unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
826 
827 	spin_lock(&pool->lock);
828 	if (!pool->ops || !pool->ops->evict || retries == 0) {
829 		spin_unlock(&pool->lock);
830 		return -EINVAL;
831 	}
832 	for (i = 0; i < retries; i++) {
833 		if (list_empty(&pool->lru)) {
834 			spin_unlock(&pool->lock);
835 			return -EINVAL;
836 		}
837 		list_for_each_prev(pos, &pool->lru) {
838 			page = list_entry(pos, struct page, lru);
839 			if (test_bit(PAGE_HEADLESS, &page->private))
840 				/* candidate found */
841 				break;
842 
843 			zhdr = page_address(page);
844 			if (!z3fold_page_trylock(zhdr))
845 				continue; /* can't evict at this point */
846 			kref_get(&zhdr->refcount);
847 			list_del_init(&zhdr->buddy);
848 			zhdr->cpu = -1;
849 			set_bit(UNDER_RECLAIM, &page->private);
850 			break;
851 		}
852 
853 		list_del_init(&page->lru);
854 		spin_unlock(&pool->lock);
855 
856 		if (!test_bit(PAGE_HEADLESS, &page->private)) {
857 			/*
858 			 * We need encode the handles before unlocking, since
859 			 * we can race with free that will set
860 			 * (first|last)_chunks to 0
861 			 */
862 			first_handle = 0;
863 			last_handle = 0;
864 			middle_handle = 0;
865 			if (zhdr->first_chunks)
866 				first_handle = encode_handle(zhdr, FIRST);
867 			if (zhdr->middle_chunks)
868 				middle_handle = encode_handle(zhdr, MIDDLE);
869 			if (zhdr->last_chunks)
870 				last_handle = encode_handle(zhdr, LAST);
871 			/*
872 			 * it's safe to unlock here because we hold a
873 			 * reference to this page
874 			 */
875 			z3fold_page_unlock(zhdr);
876 		} else {
877 			first_handle = encode_handle(zhdr, HEADLESS);
878 			last_handle = middle_handle = 0;
879 		}
880 
881 		/* Issue the eviction callback(s) */
882 		if (middle_handle) {
883 			ret = pool->ops->evict(pool, middle_handle);
884 			if (ret)
885 				goto next;
886 		}
887 		if (first_handle) {
888 			ret = pool->ops->evict(pool, first_handle);
889 			if (ret)
890 				goto next;
891 		}
892 		if (last_handle) {
893 			ret = pool->ops->evict(pool, last_handle);
894 			if (ret)
895 				goto next;
896 		}
897 next:
898 		if (test_bit(PAGE_HEADLESS, &page->private)) {
899 			if (ret == 0) {
900 				free_z3fold_page(page);
901 				return 0;
902 			}
903 			spin_lock(&pool->lock);
904 			list_add(&page->lru, &pool->lru);
905 			spin_unlock(&pool->lock);
906 		} else {
907 			z3fold_page_lock(zhdr);
908 			clear_bit(UNDER_RECLAIM, &page->private);
909 			if (kref_put(&zhdr->refcount,
910 					release_z3fold_page_locked)) {
911 				atomic64_dec(&pool->pages_nr);
912 				return 0;
913 			}
914 			/*
915 			 * if we are here, the page is still not completely
916 			 * free. Take the global pool lock then to be able
917 			 * to add it back to the lru list
918 			 */
919 			spin_lock(&pool->lock);
920 			list_add(&page->lru, &pool->lru);
921 			spin_unlock(&pool->lock);
922 			z3fold_page_unlock(zhdr);
923 		}
924 
925 		/* We started off locked to we need to lock the pool back */
926 		spin_lock(&pool->lock);
927 	}
928 	spin_unlock(&pool->lock);
929 	return -EAGAIN;
930 }
931 
932 /**
933  * z3fold_map() - maps the allocation associated with the given handle
934  * @pool:	pool in which the allocation resides
935  * @handle:	handle associated with the allocation to be mapped
936  *
937  * Extracts the buddy number from handle and constructs the pointer to the
938  * correct starting chunk within the page.
939  *
940  * Returns: a pointer to the mapped allocation
941  */
942 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
943 {
944 	struct z3fold_header *zhdr;
945 	struct page *page;
946 	void *addr;
947 	enum buddy buddy;
948 
949 	zhdr = handle_to_z3fold_header(handle);
950 	addr = zhdr;
951 	page = virt_to_page(zhdr);
952 
953 	if (test_bit(PAGE_HEADLESS, &page->private))
954 		goto out;
955 
956 	z3fold_page_lock(zhdr);
957 	buddy = handle_to_buddy(handle);
958 	switch (buddy) {
959 	case FIRST:
960 		addr += ZHDR_SIZE_ALIGNED;
961 		break;
962 	case MIDDLE:
963 		addr += zhdr->start_middle << CHUNK_SHIFT;
964 		set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
965 		break;
966 	case LAST:
967 		addr += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
968 		break;
969 	default:
970 		pr_err("unknown buddy id %d\n", buddy);
971 		WARN_ON(1);
972 		addr = NULL;
973 		break;
974 	}
975 
976 	z3fold_page_unlock(zhdr);
977 out:
978 	return addr;
979 }
980 
981 /**
982  * z3fold_unmap() - unmaps the allocation associated with the given handle
983  * @pool:	pool in which the allocation resides
984  * @handle:	handle associated with the allocation to be unmapped
985  */
986 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
987 {
988 	struct z3fold_header *zhdr;
989 	struct page *page;
990 	enum buddy buddy;
991 
992 	zhdr = handle_to_z3fold_header(handle);
993 	page = virt_to_page(zhdr);
994 
995 	if (test_bit(PAGE_HEADLESS, &page->private))
996 		return;
997 
998 	z3fold_page_lock(zhdr);
999 	buddy = handle_to_buddy(handle);
1000 	if (buddy == MIDDLE)
1001 		clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1002 	z3fold_page_unlock(zhdr);
1003 }
1004 
1005 /**
1006  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1007  * @pool:	pool whose size is being queried
1008  *
1009  * Returns: size in pages of the given pool.
1010  */
1011 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1012 {
1013 	return atomic64_read(&pool->pages_nr);
1014 }
1015 
1016 /*****************
1017  * zpool
1018  ****************/
1019 
1020 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1021 {
1022 	if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1023 		return pool->zpool_ops->evict(pool->zpool, handle);
1024 	else
1025 		return -ENOENT;
1026 }
1027 
1028 static const struct z3fold_ops z3fold_zpool_ops = {
1029 	.evict =	z3fold_zpool_evict
1030 };
1031 
1032 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1033 			       const struct zpool_ops *zpool_ops,
1034 			       struct zpool *zpool)
1035 {
1036 	struct z3fold_pool *pool;
1037 
1038 	pool = z3fold_create_pool(name, gfp,
1039 				zpool_ops ? &z3fold_zpool_ops : NULL);
1040 	if (pool) {
1041 		pool->zpool = zpool;
1042 		pool->zpool_ops = zpool_ops;
1043 	}
1044 	return pool;
1045 }
1046 
1047 static void z3fold_zpool_destroy(void *pool)
1048 {
1049 	z3fold_destroy_pool(pool);
1050 }
1051 
1052 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1053 			unsigned long *handle)
1054 {
1055 	return z3fold_alloc(pool, size, gfp, handle);
1056 }
1057 static void z3fold_zpool_free(void *pool, unsigned long handle)
1058 {
1059 	z3fold_free(pool, handle);
1060 }
1061 
1062 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1063 			unsigned int *reclaimed)
1064 {
1065 	unsigned int total = 0;
1066 	int ret = -EINVAL;
1067 
1068 	while (total < pages) {
1069 		ret = z3fold_reclaim_page(pool, 8);
1070 		if (ret < 0)
1071 			break;
1072 		total++;
1073 	}
1074 
1075 	if (reclaimed)
1076 		*reclaimed = total;
1077 
1078 	return ret;
1079 }
1080 
1081 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1082 			enum zpool_mapmode mm)
1083 {
1084 	return z3fold_map(pool, handle);
1085 }
1086 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1087 {
1088 	z3fold_unmap(pool, handle);
1089 }
1090 
1091 static u64 z3fold_zpool_total_size(void *pool)
1092 {
1093 	return z3fold_get_pool_size(pool) * PAGE_SIZE;
1094 }
1095 
1096 static struct zpool_driver z3fold_zpool_driver = {
1097 	.type =		"z3fold",
1098 	.owner =	THIS_MODULE,
1099 	.create =	z3fold_zpool_create,
1100 	.destroy =	z3fold_zpool_destroy,
1101 	.malloc =	z3fold_zpool_malloc,
1102 	.free =		z3fold_zpool_free,
1103 	.shrink =	z3fold_zpool_shrink,
1104 	.map =		z3fold_zpool_map,
1105 	.unmap =	z3fold_zpool_unmap,
1106 	.total_size =	z3fold_zpool_total_size,
1107 };
1108 
1109 MODULE_ALIAS("zpool-z3fold");
1110 
1111 static int __init init_z3fold(void)
1112 {
1113 	/* Make sure the z3fold header is not larger than the page size */
1114 	BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1115 	zpool_register_driver(&z3fold_zpool_driver);
1116 
1117 	return 0;
1118 }
1119 
1120 static void __exit exit_z3fold(void)
1121 {
1122 	zpool_unregister_driver(&z3fold_zpool_driver);
1123 }
1124 
1125 module_init(init_z3fold);
1126 module_exit(exit_z3fold);
1127 
1128 MODULE_LICENSE("GPL");
1129 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1130 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");
1131